TASP: A Time-Phased Model for Sepsis Prediction

Abstract

Background: As a lift-threatening condition, sepsis is one of the major public health issues around the world. Early prediction can improve the sepsis outcomes by prompt treatmentMethod: As part of the Physionet/Computing in Cardiology Challenge 2019, our team (FlyingBubble) proposed a Time-phAsed model for Sepsis Prediction (TASP). Realizing the fact that the incidence of sepsis is time-dependent, our model is a fusion of different frameworks along the time dimension. In the beginning stage of ICU stay, a gradient boosting tree model is used to figure out the patients with relatively high risk of sepsis. Following that, another tree model with more features is adopted to identify the risk in middle stage. If a patient stayed in ICU more than 50 hours, a deep learning framework will be used to capture the long-term relations for sepsis risk prediction in late stage. We construct proper features for each sub-models with different missing value imputation strategiesResult: The proposed model obtains a score of 0.415 by means of 10-fold cross-validation on the training set. Two simplified versions of the model respectively get scores 0.420 and 0.419 on official online test set A. And the higher one is ranked in 4th with score 0.337 on full test set.